🎠 MagicPony: Learning Articulated 3D Animals in the Wild (CVPR 2023)

Project Page | Video | Paper

Shangzhe Wu*, Ruining Li*, Tomas Jakab*, Christian Rupprecht, Andrea Vedaldi (*equal contribution)

Visual Geometry Group, University of Oxford

teaser.mp4

This method learns single-image 3D reconstruction models of articulated animal categories, from just online photo collections, without any 3D geometric supervision or template shapes.

Setup (with conda)

1. Install dependencies

conda env create -f environment.yml

or manually:

conda install -c conda-forge setuptools=59.5.0 numpy=1.23.1 matplotlib=3.5.3 opencv=4.6.0 pyyaml=6.0 tensorboard=2.10.0 trimesh=3.9.35 configargparse=1.5.3 einops=0.4.1 moviepy=1.0.1 ninja=1.10.2 imageio=2.21.1 pyopengl=3.1.6 gdown=4.5.1
pip install glfw xatlas

2. Install PyTorch

conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge

Note: The code is tested with PyTorch 1.10.0 and CUDA 11.3.

3. Install NVDiffRec dependencies

pip install git+https://github.com/NVlabs/nvdiffrast/
pip install --global-option="--no-networks" git+https://github.com/NVlabs/tiny-cuda-nn@v1.6#subdirectory=bindings/torch
imageio_download_bin freeimage

Note: The code is tested with tinycudann=1.6 and it requires GCC/G++ > 7.5 (conda's gxx also works: conda install -c conda-forge gxx_linux-64=9.4.0).

4. Install PyTorch3D (for visulaization)

conda install -c fvcore -c iopath -c conda-forge fvcore iopath
conda install -c bottler nvidiacub
conda install pytorch3d -c pytorch3d

Data

The preprocessed datasets can be downloaded using the scripts in data/, including birds, horses, giraffes, zebras and cows:

cd data
sh download_horse_combined.sh
sh download_horse_videos.sh
sh download_bird_videos.sh
sh download_giraffe_coco.sh
sh download_zebra_coco.sh
sh download_cow_coco.sh

Note: horse_combined consists of horse_videos from DOVE and additional images from Weizmann Horse Database, PASCAL, and Horse-10.

Tetrahedral Grids

Download the tetrahedral grids:

cd data/tets
sh download_tets.sh

Pretrained Models

The pretrained quadruped models can be downloaded using the scripts in results/:

cd results/quadrupeds
sh download_pretrained_horse.sh
sh download_pretrained_giraffe.sh
sh download_pretrained_zebra.sh
sh download_pretrained_cow.sh

as well as the pretrained bird model:

cd results/birds
sh download_pretrained_bird.sh

Run

Train and Testing

Check the configuration files in config/ and run, eg:

python run.py --config config/quadrupeds/train_horse.yml --gpu 0 --num_workers 4
python run.py --config config/quadrupeds/test_horse.yml --gpu 0 --num_workers 4

The test script saves some visualizations, including reconstruction images from input view, viewpoint and the resulting mesh in .obj format. For more visualizations, use the scripts/visualize_results.py as explained below.

Visualize Results

Check scripts/visualize_results.py and run, eg:

python scripts/visualize_results.py \
--input_image_dir path/to/folder/that/contains/input/images \
--config path/to/the/test/config.yml \
--checkpoint_path path/to/the/pretrained/checkpoint.pth \
--output_dir folder/to/save/the/renderings \
--render_modes input_view other_views

--input_image_dir should contain test images named as *_rgb.*.

Supported --render_modes include:

input_view: image rendered from the input viewpoint of the reconstructed textured mesh, shading map, gray shape visualization
other_views: image rendered from 12 viewpoints rotating around the vertical axis of the reconstructed textured mesh, gray shape visualization
rotation: video rendered from continuously rotating viewpoints around the vertical axis of the reconstructed textured mesh, gray shape visualization
animation (only supported for horses): two videos rendered from both a side viewpoint and continously rotating viewpoints of the reconstructed textured mesh animated by interpolating a sequence of pre-configured articulation parameters. --arti_param_dir points to ./scripts/animation_params by default which contains a sequence of pre-computed keyframe articulation parameters.
canonicalization (only supported for horses): video of the reconstructed textured mesh morphing from the input pose to a pre-configured canonical pose

Test-time Texture Finetuning

To enable test time texture finetuning, use the flag --finetune_texture, and (optionally) adjust the number of finetune iterations --finetune_iters and learning rate --finetune_lr.

For more precise texture optimization, provide instance masks in the same folder as *_mask.png. Otherwise, the background pixels might be pasted onto the object if shape predictions are not perfect aligned.

Evaluation

We evaluate our method quatitatively with the keypoint transfer metric on the PASCAL dataset, following A-CSM. To compute the metric, we first extract the mesh vertices with scripts/visualize_results.py, with a slight abuse of the script:

python scripts/visualize_results.py \
--input_image_dir path/to/preprocessed/pascal/images \
--config path/to/the/test/config.yml \
--checkpoint_path path/to/the/pretrained/checkpoint.pth \
--output_dir folder/to/save/the/output \
--evaluate_keypoint

Note we specified the --evaluate_keypoint flag.

Then, use the following to compute the keypoint transfer metric:

python evaluation/evaluate.py \
--acsm_annotations_root path/to/acsm/annotation/root \
--pascal_category horse \
--data_dir_test path/to/preprocessed/pascal/images \
--predictions_test_dir the/output/dir/from/the/previous/command \
--box_pad_frac 0.05

The code should automatically download the A-CSM annotations to the specified path, providing the user has write access to the path. If the script fails to download the annotations, they can be downloaded manually via Dropbox. The --acsm_annotations_root should be the path to the folder that contains the extracted cachedir folder from the tar file.

As A-CSM forgot to correct for their box padding, we add the 0.05 here for consistency.

TODO

Citation

@InProceedings{wu2023magicpony,
  author    = {Shangzhe Wu and Ruining Li and Tomas Jakab and Christian Rupprecht and Andrea Vedaldi},
  title     = {{MagicPony}: Learning Articulated 3D Animals in the Wild},
  booktitle = {CVPR},
  year      = {2023}
}

elliottwu / MagicPony